Source
Olah, Cammarata, Schubert, Goh, Petrov and Carter (2020), Zoom In: An Introduction to Circuits, Distill, LIT-1543, read in full in NOTE-1182: Claims 1 and 2, the glossary, and Interpretability as a Natural Science.
What is asserted
That a trained network “consist[s] of meaningful, understandable features”, directions in activation space, usually neurons, and that the weights between features in adjacent layers are “meaningful algorithms corresponding to facts about the world” that can be read off. The essay offers these as empirical claims and as norms for how to study networks.
What the record holds toward it
- Curve detectors in InceptionV1’s mixed3b, argued seven ways, including a hand-set reimplementation that mimics the originals.
- Three circuits: curve detectors built from earlier curve detectors along the tangent, with weights that rotate with orientation; a four-layer oriented dog head circuit; and a car feature spread over dog neurons.
- Its own counter-case: polysemantic neurons such as 4e:55 (cat faces, car fronts, cat legs), “a major challenge for the circuits agenda”.
Why it stays Proposed
The essay calls its claims speculative and defers the evidence. The examples were chosen, from one model, with no denominator, and the counter-case is acknowledged and unresolved. The transformer work that follows (LIT-1534) reports it “had a lot more success” with attention heads than with neurons, and that MLP neurons have been “much harder to develop hypotheses for”.
What this does not say
- Not that every neuron is a feature: polysemantic neurons are granted.
- Not that features are unique directions: a direction is a linear combination of neurons, and the essay does not ask whether another basis would serve as well (the question of THEORY-802).
- Not anything about how features or circuits form in training.
What would confirm or reject it
The promote_when above.
Note of 2026-10-27: Curve Detectors, read
Cammarata et al.’s Curve Detectors (LIT-1575, NOTE-1198), the first of the articles Zoom In defers its evidence to, was registered and read on 2026-10-27 at the owner’s request. Of Zoom In’s seven arguments for a curve detector, it supplies four for the mixed3b family (feature visualization, dataset examples, synthetic curves and joint tuning under rotation, with attribution and blind labelling added), measured as THEORY-1028. The other three (the implementing circuit, the downstream use and the hand-written reimplementation) are deferred again, to Curve Circuits, which the record does not hold. The authors’ full claim of causality, generality and purity is THEORY-1022, Proposed.
The article gives no count over a layer: the family is ten of mixed3b’s 480 units, about 2%. The promote_when’s “fraction of units that meet” a stated criterion is therefore not supplied, and the account stays Proposed for the reason it already gives: chosen examples, no denominator. The likeliest source of such a count is the thread’s An Overview of Early Vision in InceptionV1, which sorts the early layers’ units into families; it is not registered.
Note of 2026-10-27 (later): Early Vision and Curve Circuits, read
The note above says that An Overview of Early Vision in InceptionV1 “is not registered” and that Curve Circuits is not held. Both were registered and read on 2026-10-27, at the owner’s request (“file the next readings”): Early Vision as LIT-1624 (NOTE-1206) and Curve Circuits as LIT-1629 (NOTE-1207).
- A denominator, not the criterion (THEORY-1042). Every one of the 1,056 units of conv2d0 to mixed3b is in exactly one of 77 groups (3, 8, 14, 18 and 34 per layer). 391 units (37%) are in groups whose own captions call them catch-all, broad and not well defined, low-confidence, unstudied or not understood: 14%, 48%, 7%, 48% and 45% by layer. The fourteen mixed3b groups the article presents as well understood hold 118 of 480 units (24.6%), and the curve family is 10 (2.1%). The groups are sorted by eye, often after “a few minutes”, so this is the fraction the promote_when asks for without the “stated criterion” it asks for. The authors’ claim that such families are real and recur across models is THEORY-1040, Proposed.
- The seven arguments, for one family (THEORY-1038). Curve Circuits supplies the three Curve Detectors deferred: the implementing circuit, a downstream use (described, not measured) and the hand-written reimplementation, compared by figures only. With THEORY-1028, all seven of Zoom In’s arguments are now held for the mixed3b curve family. All but one are qualitative. The exception is the echo edit: restricting the curve units’ input to mixed3a’s early curves with the negative weights removed brings back the 180° echo, and keeping them does not. That is one weight-level prediction tested by editing, of the kind the promote_when names. The authors’ claim that the algorithm is simple and can be read off the weights is THEORY-1047, Proposed.
Why it stays Proposed. One family in one model, one confirmed edit, and no layer-wide criterion. Early Vision’s census says how much of the early layers is sorted, not how much meets Zoom In’s seven-argument standard; by that standard only the curve family and the curve circuit’s inputs do. The status is unchanged.